Research On Base Station Network Traffic Prediction Method Based On Distributed Machine Learning

With the gradual popularization of 5G and the continuous development of network applications,people’s demand for mobile network traffic is also increasing.Base station is to provide mobile network infrastructure service,accurate,fast,base station network traffic prediction is the base station operations staff distribution network resources,ensure the normal operation of mobile network in a timely manner to upgrade network foundation,at the same time page is mobile Internet users reasonable planning time online,to avoid network congestion get a better browsing experience.Therefore,the accurate and rapid prediction of the base station network flow is not only crucial to the stability and normal operation of the telecommunication system,but also has a significant impact on the production and life of the whole society.Therefore,efficient and accurate prediction of base station network traffic is an important task.Recurrent neural network has excellent performance in base station network traffic prediction,but it needs a lot of data to train the model.For the sake of privacy and security,major operators are not willing to share their own base station network traffic data,which makes it difficult to accurately predict base station network traffic.In addition,the transmission of base station network traffic data to the central server will generate huge network resource overhead,which will further increase the possibility of network congestion.To solve these problems,this paper combines federated learning with deep learning to design and implement two base station network traffic prediction methods to solve the problem of data leakage,and at the same time ensure the accuracy of prediction model and improve the speed of model training.The main contents of this thesis are as follows:（1）Firstly,the importance of accurate base station network traffic prediction is described,and the existing base station network traffic prediction methods are summarized and summarized.The advantages and disadvantages of these methods are pointed out,and then the federated learning technology is introduced to solve the problems of insufficient data amount in the training process of recurrent neural network model,huge resource cost and data leakage in the process of data uploading.Then,federated learning theory,recurrent neural network theory and Py Torch deep learning framework are studied,which provides a basis for the construction and training of recurrent neural network model.（2）Then propose a method of LSTM base station network traffic prediction based on federated learning.To training data can be directly used to model and forecast,and raise the convergence speed and accuracy,model training to modify the data set of outliers and to fill the missing value,at the same time,in order to avoid different operators or base station at the time of collecting traffic data using different criteria,lead to an order of magnitude difference in network traffic data influence on the model of training,We standardize the data.Then,two common aggregation algorithms in federated learning are introduced in detail.The system design of LSTM base station network traffic prediction method based on federated learning is introduced in detail,including LSTM model federated learning framework,construction of LSTM model,LSTM model local training process in base station and LSTM model aggregation process in central server.Finally on real data sets,the centralized training method and the training methods for experiments,and analysis the result of the experiment,the experimental results show that the method of learning to get LSTM the precision of the model close to the LSTM models obtained from the method of centralized training,but the learning communication efficiency significantly higher than that of the centralized approach.（3）In practice,there are many factors that affect the base station network traffic.The LSTM neural network model mentioned above only uses historical network traffic data to train the model,while the DeepAR neural network model can learn other features related to the base station network traffic sequence to further improve the model accuracy.Such as hour cycle,daily cycle,weekly cycle and holidays and other time characteristics.Then,the system design of DeepAR base station network traffic prediction method based on federated learning is introduced in detail,including the framework of federated learning algorithm of DeepAR model,construction of DeepAR model,local training of DeepAR model and DeepAR federated aggregation process.Finally,LSTM model and DeepAR model are trained and compared on real data sets.The comparison results show that the PREDICTION effect of DeepAR model is better than that of LSTM model on base station network traffic.